Mapp: A model-agnostic privacy-preserving framework for two-party graph neural network inference.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Privacy

Background:

• Graph neural networks (GNNs) offer significant advancements in healthcare and finance.
• Deploying GNNs raises privacy concerns for sensitive client data and proprietary models.
• Existing secure inference protocols struggle with GNNs' unique operational demands and architecture diversity.

Purpose of the Study:

• To develop a model-agnostic privacy-preserving framework for GNN inference.
• To address the limitations of current secure inference protocols in handling GNNs.
• To enable efficient and secure deployment of diverse GNN architectures.

Main Methods:

• Introduced MAPP, a novel privacy-preserving framework for GNNs.
• Utilized knowledge distillation to train lightweight proxy GNNs adaptable to various architectures.
• Developed graph-aware protocol optimizations, including sparsity exploitation and offline computation for encrypted multiplications.

Main Results:

• 79% of proxy models achieved target accuracy, with some improving by 5 percentage points.
• Achieved significant reductions in online latency (8.8-14.2x) and communication (4.8-7.6x).
• Demonstrated the framework's effectiveness across 7 datasets and 4 GNN architectures.

Conclusions:

• MAPP provides an efficient and broadly applicable solution for privacy-preserving GNN inference.
• The model-agnostic approach and optimized protocol reduce complexity and redesign efforts.
• Enables secure deployment of GNNs in sensitive domains without compromising performance.